search

universome / stylegan-v

[CVPR 2022] StyleGAN-V: A Continuous Video Generator with the Price, Image Quality and Perks of StyleGAN2
https://universome.github.io/stylegan-v
348 stars 35 forks source link

Question about FVD calculation #30

Closed hse1032 closed 1 year ago

hse1032 commented 1 year ago

Thanks you for sharing this great codebase!

I have some question about the FVD caculation in this repository.

As noted in paper, you use (train + test) split for UCF-101 dataset and (train) split for other datasets.

Then, when you calculating the real statistics of FVD, did you also sample real data from (train + test) split for UCF-101 dataset and (train) split for other datasets?

Thanks,

universome commented 1 year ago

Hi, @hse1032 ! The real statistics for FVD were computed on the set of images that was used to train the model. For UCF, we used the train+test split to compute the real statistics for FVD. For other datasets, we used the train split (e.g., we discarded the val/test splits for them if they were available).

hse1032 commented 1 year ago

Thank you for prompt reply!